794 pop-music chorus clips, annotated by 457 listeners along the valence-arousal plane, with simultaneous electrodermal-activity (EDA / skin conductance) recordings. An open dataset for music emotion recognition, published at ICMR 2018 (Yokohama).
SOURCE From Billboard Hot 100, iTunes Top 100 and UK Top 40 Singles charts (2016–2017), an initial pool of 1,000 popular songs was deduplicated to 794. For each song, the chorus segment was manually selected by music students, then annotated by ≥10 listeners (266 Chinese non-music students + 44 music majors + 47 native English speakers). Skin conductance was recorded simultaneously at 50 Hz.
3 charts contributed the original 1,000-song pool. Overlap removed → 794 unique chorus clips.
Pop-music skews toward Q1 (high V, high A) — bright + energetic. Q4 (calm) is rare.
Drake leads the pack — typical of late-2010s pop-chart dominance.
PROTOCOL Each clip is a point on the (V, A) plane, where V ∈ [0, 1] is perceived pleasantness and A ∈ [0, 1] is perceived energy. Background heatmap shows density (20×20 bins). Replicates Fig. 3 of the paper. Click a point to inspect.
MODALITIES For each quadrant, one chorus clip with the most extreme V/A values is shown. Press play; the playhead slides along the dynamic VA curve (2 Hz) and the multi-subject EDA traces (50 Hz, downsampled, z-scored).
SOURCES Lyrics scraped as .lrc files (timestamps stripped); comments crawled from Netease Cloud Music (中文) and SoundCloud (English). Word frequencies below — note the difference in tone between the two listener communities.
中文评论偏抒情，英文评论偏直白：网易云常见 "感觉 / 喜欢 / 听到"，SoundCloud 常见 "follow / love / fire / song".
SETUP Original paper: 6373-d ComParE features → MLR / SVR → static V/A regression. Reproduced here in compressed form: PCA-60 features → RBF-SVR (C=1.0). Test split below; r values closely match the paper's baseline.
| task | model | RMSE | r |
|---|---|---|---|
| static V | MLR | 0.136 | 0.546 |
| SVR | 0.124 | 0.638 | |
| static A | MLR | 0.111 | 0.719 |
| SVR | 0.102 | 0.764 |